Alibaba made Qwen3.8-Max generally available on August 3, 2026.
This is a practical rundown for developers: what the model is, what it costs, how to call it, and where the claims still need a pinch of salt.
The one-line version
Qwen3.8-Max is a 2.4-trillion-parameter Mixture-of-Experts model with a 1M-token context window, native text/image/video input, OpenAI-compatible API, and pricing of $2 in / $6 out per million tokens. Open weights are promised for next week.
First, the naming confusion
The version number trips people up, so let's clear it up:
- Qwen3 is the open-weight model family you've probably used (Qwen3-32B, Qwen3-235B-A22B, etc.)
- Qwen3.5 / 3.6 / 3.7 / 3.8 are successive flagship generations, not point releases of Qwen3
- "Max" is the top tier of each generation
So Qwen3.8-Max is not "Qwen 3, version 8." It is the newest flagship, and it succeeds Qwen3.7-Max from May 2026. Qwen describes it as their most capable model to date, and the first open-weight model at Max scale.
Also note there is a Qwen3.8-Max-Preview (July 19) and now Qwen3.8-Max (August 3, GA). If you're reading a blog post from July, it's about the preview, and half its "not disclosed yet" list has since been answered.
The specs that matter
| Spec | Value |
|---|---|
| Total parameters | 2.4 trillion (sparse MoE) |
| Active parameters per token | ~95 billion (reported) |
| Context window | 1,000,000 tokens |
| Max input | 991K tokens (983K with thinking on) |
| Max output | 131K tokens |
| Max reasoning budget | 262K tokens |
| Input modalities | Text, image, video |
| Output | Text |
| Rate limits | 2M tokens/min, 15K requests/min |
| Model ID | qwen3.8-max |
Why "95B active" is the number to care about
This is the part worth understanding properly, because "2.4 trillion parameters" is close to meaningless on its own.
A sparse Mixture-of-Experts model is not one giant network. Each layer holds many specialist sub-networks ("experts"), and a router picks a small handful for each token. The rest sit idle for that token.
Qwen's own smaller models make the pattern obvious from the naming: Qwen3-235B-A22B carries 235B total parameters but activates 22B per token, and Qwen3-30B-A3B activates roughly 3B.
So:
- Total parameters ≈ how much the model knows (and how much memory you'd need to host it)
- Active parameters ≈ how much compute each token actually costs
At 2.4T total / ~95B active, roughly 4% of the network fires per token. That's why Alibaba can sell it at $2/M input rather than something ruinous.
One honesty note: this figure needs a small asterisk. MarkTechPost's launch coverage stated Alibaba had not disclosed the activated-parameter count, while benchmark trackers report that Qwen's own August 3 release post lists 2.4T total with 95B active. Other coverage advises treating the number as reported rather than confirmed until Alibaba publishes a model card. Use it for rough cost intuition, not for capacity planning.
And to be clear: 95B active does not mean you can serve this on a 95B-sized box. A serving system still needs fast access to the full expert pool, plus attention state, routing machinery, multimodal components and runtime buffers. The whole 2.4T checkpoint has to be resident somewhere.
Pricing, and the one trick that dominates your bill
| Item | Price per 1M tokens |
|---|---|
| Input | $2.00 |
| Output | $6.00 |
| Input (implicit cache read) | $0.25 |
| Explicit cache creation | $2.50 |
| Explicit cache read | $0.17 |
Here's the thing to internalize: cached input is eight times cheaper than fresh input, which means prefix stability drives your cost more than prompt length does.
That single sentence should change how you architect against this model. Concretely, imagine an agent loop with a 200K-token stable prefix (system prompt, tool schemas, codebase context) running 50 turns:
No caching: 50 × 0.2M × $2.00 = $20.00
Explicit cache: (0.2M × $2.50) + (50 × 0.2M × $0.17) = $2.20
Roughly a 9x difference, from nothing but keeping your prefix byte-stable. Practical implications:
- Put everything static at the front of your prompt, and everything variable at the end
- Don't inject timestamps, request IDs or shuffled context into your system prompt
- If you're rebuilding the prefix per request, you're paying 8x for no reason
Compare against the field: Kimi K3 runs $3.00 input / $15.00 output per million tokens, so Qwen3.8-Max undercuts it meaningfully, particularly on output.
Calling the API
Integration is deliberately boring, which is the point. The hosted API is OpenAI- and DashScope-compatible, so integration is a base-URL and model-ID change. It also supports the Anthropic protocol, so tools such as Cursor, Cline, Codex and Claude Code can point at it through existing integrations.
Here's the multimodal example straight from the model page, using the DashScope SDK:
import os
import dashscope
dashscope.base_http_api_url = "https://dashscope-intl.aliyuncs.com/api/v1"
messages = [
{
"role": "user",
"content": [
{"image": "https://example.com/your-image.jpeg"},
{"text": "What is depicted in the image?"}
]
}
]
response = dashscope.MultiModalConversation.call(
api_key=os.getenv('DASHSCOPE_API_KEY'),
model='qwen3.8-max',
messages=messages
)
print(response.output.choices[0].message.content[0]["text"])
If you're already on the OpenAI SDK, you point base_url at Alibaba's compatible-mode endpoint and change the model string to qwen3.8-max. Grab the exact base URL from docs.qwencloud.com rather than guessing, since it differs between the international and China-domestic deployments.
Watch out for the two regional endpoints. Availability is split between an international deployment (qwencloud.com) and a China-domestic one (platform.qianwenai.com), each requiring separate registration and billing. Keys are not interchangeable.
It's also on third-party gateways already. Vercel's AI Gateway added it as alibaba/qwen3.8-max on August 2 at provider pricing with no markup.
Supported features
Function calling, structured outputs, batches, prefix completion and fine-tuning are all supported. Five built-in tools ship on the Responses API: code_interpreter, web_search, web_extractor, t2i_search and i2i_search.
The built-in tools are worth a look before you hand-roll your own. If you're currently maintaining a custom web-search tool wrapper, that's now a server-side flag.
Benchmarks: the honest read
Unlike the July preview, GA came with actual numbers. Alibaba published a benchmark table with the formal launch showing 86.6 on Terminal-Bench 2.1, 67.7 on SWE-bench Pro, and 92.6 on GPQA Diamond, with the strongest gains in multimodal and agentic categories rather than general reasoning.
Filling in the competitive picture:
| Benchmark | Qwen3.8-Max | Claude Fable 5 | Notes |
|---|---|---|---|
| Terminal-Bench 2.1 | 86.6 | 84.6 | GPT-5.6 Sol (max) leads at 88.8 |
| SWE-bench Pro | 67.7 | 80.0 | Fable 5 ahead |
| FrontierSWE | 73.5 | 88.8 | Fable 5 ahead |
| GPQA Diamond | 92.6 | — | Up marginally from 3.7-Max's 92.4 |
| PaperBench | 93.0 | — | Qwen leads |
| IFBench | 82.8 | — | Qwen leads |
It also tops most vision rows, including OSWorld-Verified at 86.1, Parametric CAD Bench at 91.5, and OmniDocBench 1.5 at 92.1. Against its own predecessor the jump is large: DeepSWE 1.1 moves from 21.6 to 56.6, FrontierSWE from 40.7 to 73.5, and JobBench from 31.3 to 53.4.
Two caveats that belong in any fair reading, both flagged by MarkTechPost: the multimodal table benchmarks against Qwen3.7-Plus rather than Qwen3.7-Max, which flatters the generational delta; and Alibaba's own RL scaling curve peaks at 0.725 near 4,000 training environments, then declines to 0.719 and 0.689.
Open weights: read the fine print
Alibaba confirmed open weights ship next week, along with a second checkpoint, Qwen3.8-27B. Releases are expected the week of August 10 via Alibaba Cloud Model Studio. This marks Alibaba's return to open-sourcing its top-tier models after keeping several recent flagships proprietary earlier this year.
What's still missing
Being straight about the gaps:
- No model card. The GA launch still did not provide an official training and safety model card. No training data disclosure, no safety evaluation methodology.
- No published license. Until the weights land, there's nothing to review on commercial-use terms.
- Active parameter count is reported, not officially confirmed in a spec sheet.
- No independent Artificial Analysis score yet.
Should you use it?
Try it now if: you're doing multimodal work (documents, video indexing, screenshots, UI automation), you want long-context agent runs at a fraction of Western frontier pricing, or you have an existing OpenAI/Anthropic-protocol setup where testing costs you a base-URL change and an afternoon.
References:
- Qwen3.8-Max model page (QwenCloud) — canonical specs, pricing, rate limits
Running this in production yet? I'd be interested in real latency and tokens-per-second numbers, since Alibaba hasn't published throughput figures.
Top comments (0)